Muse Spark (Meta AI)

Muse Spark is a proprietary family of multimodal large language models, reasoning models, and agent-oriented foundation models developed by Meta Superintelligence Labs (MSL), an artificial-intelligence organization within Meta Platforms. The original Muse Spark model was announced on 8 April 2026 as the first model in Meta's new Muse family and as the first production result of a ground-up reconstruction of the company's model-training stack.

Muse Spark was designed as a natively multimodal reasoning system supporting visual understanding, tool use, test-time reasoning, computer interaction, and parallel orchestration of multiple agents. It initially powered the Meta AI application and meta.ai, followed by staged integration into Meta's social applications and AI glasses.

As of 10 July 2026, the latest publicly announced model in the branch is Muse Spark 1.1, released on 9 July 2026. Meta describes it as a multimodal reasoning model optimized for agentic workloads, computer use, software engineering, long-context operation, and multimodal understanding. It is available in Thinking mode through Meta AI and through the public preview of the Meta Model API.

The hosted muse-spark-1.1 endpoint accepts text, images, video, and PDF documents and produces text. Its documented context window is 1,048,576 tokens, shared between the request and the requested output. The video interface can also inspect audio embedded in a video file, but the model catalogue does not list independent raw-audio input as a separate Muse Spark 1.1 API modality.

Meta has not published Muse Spark's parameter count, layer count, hidden-state size, tokenizer specification, expert-routing topology, complete training-token count, or full training-data inventory. No downloadable Muse Spark or Muse Spark 1.1 weight release, self-hosting package, or open model licence had been announced by 10 July 2026. Consequently, claims that Muse Spark uses a particular dense, sparse, or Mixture-of-Experts architecture remain speculative unless Meta publishes a technical specification.

Muse Spark is distinct from Muse Image and Muse Video, separate media-generation models in the wider Muse family. Muse Spark 1.1 itself returns text through the public API, although Meta AI can combine it with search, code execution, user-interface tools, and Muse media models to create images, videos, websites, animations, and other artifacts.

History and background

Development within Meta Superintelligence Labs

Meta presented Muse Spark as a departure from simply extending the published Llama sequence. According to the company, Meta Superintelligence Labs spent approximately nine months rebuilding the underlying development stack, including model architecture research, optimization, data curation, reinforcement-learning infrastructure, evaluation, and deployment systems.

The first Muse Spark release was deliberately described as “small and fast” relative to the larger systems then under development. Meta positioned it as the first point on a scaling ladder: an initial model intended to validate a new technical stack before training substantially larger successors.

Meta also linked the programme to continuing investment in data-centre and accelerator infrastructure, including the Hyperion data-centre project. Public launch materials did not disclose the exact training-cluster size, accelerator model, total training FLOPs, or duration of the final Muse Spark training run.

Relationship to Llama

Muse Spark and Llama are separate Meta model brands. Llama has historically been distributed as an open-weight family with downloadable checkpoints and model-specific community licences. Muse Spark was launched as a hosted proprietary system integrated into Meta products, followed by access through a metered cloud API.

Meta compared the efficiency of the new Muse Spark pre-training recipe with Llama 4 Maverick. The company reported that small-scale scaling-law experiments indicated that its rebuilt stack could reach the same evaluated capability level with more than ten times less training computation than the earlier recipe used for Llama 4 Maverick. This is a developer-reported scaling-law comparison, not an independent measurement of the complete production models. It also does not imply that Muse Spark has one-tenth of Llama 4 Maverick's parameters or inference cost.

Muse Spark should not therefore be described as “Llama 5” or as an open-weight successor to Llama 4. Its architecture, licence, distribution model, product role, and level of public technical disclosure differ.

Initial launch

Muse Spark was announced on 8 April 2026. The initial release emphasized four capability groups:

• multimodal perception and visual reasoning;
• mathematical, scientific, and health-related reasoning;
• tool use and interactive artifact creation;
• multi-agent test-time computation.

The model became available through the Meta AI application and meta.ai. Meta simultaneously opened a private API preview to selected partners. At launch, the company acknowledged relative weaknesses in long-horizon agent execution and software-engineering workflows, both of which became major targets of the 1.1 update.

Contemplating mode

The April launch also introduced Contemplating mode, an inference-time reasoning configuration in which several agents explore a problem in parallel before their work is synthesized. Meta compared the approach with the high-compute or “deep thinking” modes of other frontier-model providers.

Meta's launch evaluation reported 58% on Humanity's Last Exam and 38% on FrontierScience Research for Contemplating mode. The configuration was rolled out gradually within Meta AI rather than announced as a separately downloadable checkpoint.

A separate safety report published on 3 June 2026 described Muse Spark Contemplating as a deep-reasoning model built upon the Muse Spark foundation and using multi-agent orchestration at inference time. The report did not disclose a separate parameter count or demonstrate that Contemplating used an independently pretrained backbone.

Product expansion

Meta expanded the Muse Spark-powered Meta AI product during April and May 2026. Announced features included:

• faster spoken conversations in the Meta AI application;
• live camera interactions;
• multimodal assistance based on photographs and the device camera;
• shopping search combining Facebook Marketplace and external internet results;
• recommendations drawing on public content from Meta services;
• staged deployment to WhatsApp, Instagram, Facebook, Messenger, and Threads;
• deployment to Ray-Ban Meta and Oakley Meta glasses in the United States and Canada.

On 23 June 2026, Meta announced a new Meta Glasses product line that launched with Muse Spark-powered Meta AI from its first release. The company also stated that Muse Spark was already available on supported Ray-Ban Meta and Oakley Meta glasses in the United States and Canada.

Product availability differed by country, device, application version, and account. A capability advertised for the consumer Meta AI system should not automatically be attributed to the standalone Muse Spark API, because voice processing, search, media generation, recommendation retrieval, memory, and graphical interfaces can be supplied by separate product components.

Muse Image and Muse Video

On 7 July 2026, Meta released Muse Image and previewed Muse Video. These systems share the Muse brand but serve different output modalities from Muse Spark.

Muse Image is an agentic image-generation and editing model that can invoke code and search tools, iteratively refine its output, and coordinate with Muse Spark. Muse Video is a separate video-generation system with native audio support. Meta demonstrated joint workflows in which Muse Spark plans an artifact while Muse Image generates visual components—for example, animated GIFs, illustrated websites, and interactive games.

This integration explains why the Meta AI product may appear to “generate images with Muse Spark” even though the public Muse Spark 1.1 endpoint is documented as a text-output model.

Muse Spark 1.1

Muse Spark 1.1 was released on 9 July 2026, approximately three months after the first public model. Its principal areas of improvement were:

• general-purpose tool use;
• orchestration of main agents and sub-agents;
• operation of graphical computer interfaces;
• adaptation between graphical actions and generated automation scripts;
• repository-level software engineering;
• visual debugging;
• long-context management and compaction;
• image, video, document, and embedded-audio understanding.

The model became available through Thinking mode in Meta AI and through the public preview of the Meta Model API. Meta stated that more capable Muse models remained in training.

Release chronology

Date	Model, mode, or related system	Main characteristics	Availability
8 April 2026	Muse Spark	First Muse-family model; native multimodal reasoning, visual reasoning, tool use, health applications, and multi-agent orchestration	Meta AI application, meta.ai, and private API preview
8 April 2026	Muse Spark Contemplating	High-compute reasoning configuration using parallel agents and result synthesis	Gradual rollout in Meta AI; safety report published 3 June 2026
12 May 2026	Expanded Meta AI rollout	Faster voice, live camera interaction, shopping mode, and staged expansion to social applications and AI glasses	Region- and device-dependent consumer rollout
7 July 2026	Muse Image	Separate agentic image-generation and editing model that can coordinate with Muse Spark	Meta AI, meta.ai, Instagram Stories in the United States, and limited WhatsApp markets at launch
7 July 2026	Muse Video	Separate video-generation model with native audio support	Preview; broader creator and Meta AI availability announced as forthcoming
9 July 2026	Muse Spark 1.1	Improved agents, computer use, coding, multimodal understanding, and active management of a one-million-token context	Meta AI Thinking mode and Meta Model API public preview

Technical foundations

Limits of public architectural information

Meta describes Muse Spark as a large, natively multimodal reasoning model but has not published an architecture paper comparable to the detailed reports released for some earlier Meta model families. Public materials do not specify:

• the number of parameters;
• whether the core network is dense or sparsely activated;
• the number of Transformer layers;
• the dimensions of hidden states or feed-forward layers;
• the number or type of attention heads;
• the tokenizer or vocabulary size;
• the number of visual or video tokens used for a given input;
• the internal routing mechanism, if any;
• the precision formats used during training;
• the complete optimizer and learning-rate schedule;
• the full training-token or image-count budget.

The equations in this section are therefore conceptual descriptions of documented model behaviour. They are not a reconstruction of Meta's undisclosed implementation.

Autoregressive generation

Like other generative language models, Muse Spark can be represented conceptually as assigning a conditional probability to an output token sequence:

$P_{\theta }(y_1,\dots ,y_T\mid c)=\prod _{t=1}^T{P}_{\theta }(y_t\mid y_{<t},c),$

where $c$ denotes the combined input context, potentially including text, visual representations, document content, video, tool results, instructions, and retained state.

This factorization describes the generation interface but does not reveal whether the underlying implementation is a conventional Transformer, a hybrid sequence model, a sparse architecture, or another design.

Native multimodality

The initial Muse Spark release was described as natively multimodal rather than as a text model with a separately attached product-level image-captioning service. A generic representation of multimodal conditioning is:

$\mathbf{𝐙}=\mathrm{Fuse}(E_{\mathrm{t}\mathrm{e}\mathrm{x}\mathrm{t}}(X),E_{\mathrm{i}\mathrm{m}\mathrm{a}\mathrm{g}\mathrm{e}}(I),E_{\mathrm{v}\mathrm{i}\mathrm{d}\mathrm{e}\mathrm{o}}(V),E_{\mathrm{d}\mathrm{o}\mathrm{c}\mathrm{u}\mathrm{m}\mathrm{e}\mathrm{n}\mathrm{t}}(D)),$

where:

• $E_{\mathrm{t}\mathrm{e}\mathrm{x}\mathrm{t}}$ maps text to internal representations;
• $E_{\mathrm{i}\mathrm{m}\mathrm{a}\mathrm{g}\mathrm{e}}$ maps images or image regions;
• $E_{\mathrm{v}\mathrm{i}\mathrm{d}\mathrm{e}\mathrm{o}}$ represents temporal visual information and, in the API workflow, embedded audio;
• $E_{\mathrm{d}\mathrm{o}\mathrm{c}\mathrm{u}\mathrm{m}\mathrm{e}\mathrm{n}\mathrm{t}}$ represents parsed or rendered document content;
• $\mathrm{Fuse}$ denotes an undisclosed fusion process.

Meta has not stated whether these encoders are fully separate, partly shared, or integrated into one tokenizer. A PDF is also not necessarily a primitive neural modality: the API may extract text, render pages, identify images, or combine these operations before model inference.

Muse Spark 1.1 can use visual information during an agent trajectory rather than only answering a one-time image question. Meta demonstrated workflows in which the model:

• observes a graphical interface;
• takes screenshots;
• identifies a visual failure;
• traces that failure to source code;
• edits the program;
• reruns or renders the result;
• visually checks whether the repair succeeded.

Reasoning effort and test-time computation

Muse Spark is a reasoning model whose inference budget can be varied. The Model API exposes a reasoning_effort control. The documentation describes levels including none, minimal, low, medium, and high; Meta's own Muse Spark 1.1 capability evaluations used an additional xhigh configuration.

Higher reasoning effort can increase:

• time before the first visible output token;
• the number of internal or exposed reasoning tokens;
• total inference cost;
• the probability of solving some difficult tasks;
• the opportunity for self-correction.

It does not guarantee a correct answer. Long reasoning can also reinforce a false premise, overcomplicate a simple task, or spend computation on irrelevant branches.

A simplified test-time objective consistent with Meta's description of token-efficient reasoning can be written as:

$J(\theta )={\mathbb{𝔼}}_{y\sim {\pi }_{\theta }(\cdot \mid x)}[R(x,y)-\lambda L_{\mathrm{r}\mathrm{e}\mathrm{a}\mathrm{s}\mathrm{o}\mathrm{n}}(y)],$

where $R(x,y)$ rewards successful solutions, $L_{\mathrm{r}\mathrm{e}\mathrm{a}\mathrm{s}\mathrm{o}\mathrm{n}}$ measures reasoning length or cost, and $\lambda$ controls the efficiency penalty. Meta has not published this as its exact production loss; it is a simplified abstraction of the stated training principle.

Thought compression

Meta reported that applying a penalty to thinking time produced a non-monotonic transition on selected evaluations such as AIME. The model first improved by generating longer reasoning, then learned to compress that reasoning into fewer tokens, and subsequently resumed extending solutions as additional training improved performance.

The company called this behaviour thought compression. It should not be interpreted as evidence that every short answer contains the same computation as a longer answer. The observation was reported for selected training and evaluation settings.

Agent trajectories and tool use

An agentic Muse Spark deployment can be represented as a trajectory:

$\tau =(o_0,a_1,o_1,a_2,o_2,\dots ,a_T,o_T),$

where $o_t$ is an observation and $a_t$ is an action such as:

• emitting text;
• calling a search function;
• querying a database;
• invoking a Model Context Protocol server;
• running code;
• editing a file;
• clicking or typing in a graphical interface;
• delegating a task to a sub-agent.

The model's policy may be written conceptually as:

$a_t\sim {\pi }_{\theta }(a\mid o_{\le t},a_{<t},m_t,g),$

where $m_t$ is retained or compressed memory and $g$ is the current goal.

A tool result is external evidence, not a parameter update. Muse Spark does not become permanently retrained whenever it searches the web or uses an application.

The Meta Model API supports function or tool calling, including multi-turn workflows in which tool results are returned to the model. Meta recommends the Responses API for stateful agent applications and interleaved reasoning with tools.

Multi-agent orchestration

Muse Spark's multi-agent system decomposes a task into parallel branches. In the 1.1 release description, a main agent can gather context, construct a plan, create or invoke sub-agents, delegate work, and synthesize their findings. A sub-agent is trained to follow its assigned role, use the tools available to it, and return control when escalation is needed.

For $n$ independent subtasks, sequential execution has approximate wall-clock time:

$T_{\mathrm{s}\mathrm{e}\mathrm{q}\mathrm{u}\mathrm{e}\mathrm{n}\mathrm{t}\mathrm{i}\mathrm{a}\mathrm{l}}\approx T_{\mathrm{p}\mathrm{l}\mathrm{a}\mathrm{n}}+\sum _{i=1}^n{T}_i+T_{\mathrm{s}\mathrm{y}\mathrm{n}\mathrm{t}\mathrm{h}\mathrm{e}\mathrm{s}\mathrm{i}\mathrm{s}}.$

Ideal parallel execution instead approaches:

$T_{\mathrm{p}\mathrm{a}\mathrm{r}\mathrm{a}\mathrm{l}\mathrm{l}\mathrm{e}\mathrm{l}}\approx T_{\mathrm{p}\mathrm{l}\mathrm{a}\mathrm{n}}+\underset{i}{\max }{T}_i+T_{\mathrm{s}\mathrm{y}\mathrm{n}\mathrm{t}\mathrm{h}\mathrm{e}\mathrm{s}\mathrm{i}\mathrm{s}}.$

Real systems also incur scheduling, duplicated-search, communication, context-transfer, and synthesis costs. Parallelism is therefore most useful when subtasks are sufficiently independent. It can offer little benefit—or reduce quality—when every step depends tightly on the previous one.

Muse Spark 1.1 was explicitly trained for both sides of this relationship: operating as the orchestrator and operating as a constrained sub-agent.

Active context management

Muse Spark 1.1 supports a one-million-token context and was trained to manage that context actively. Meta describes three related behaviours:

• remembering earlier actions;
• retrieving information from much earlier in a workflow;
• compacting history while preserving steps judged critical for future work.

A conceptual memory update can be expressed as:

$m_t=C_{\theta }(m_{t-1},o_t,a_t,g),$

where $C_{\theta }$ is a learned or scaffold-assisted compaction function.

Context compaction is lossy unless the retained representation reproduces every detail. The system must decide what is important before it knows all future needs. An incorrectly discarded constraint, source, error message, or user preference can cause later failure.

The one-million-token figure is also a combined budget: request tokens and requested output tokens share the maximum. A request near the limit cannot necessarily produce an equally large answer.

Computer use

Muse Spark 1.1 was trained for workflows spanning multiple graphical applications. Meta states that the model can choose among three broad execution strategies:

• direct graphical interaction, such as clicking, typing, scrolling, or selecting;
• writing and running scripts when automation is more efficient;
• batching several actions when they can safely be planned together.

This differs from a model that must emit one isolated mouse action after every screenshot. Script generation can reduce the number of interface steps, while direct interaction remains useful when an application lacks a suitable programmatic interface.

Computer-use agents require an external runtime that captures the screen, transmits observations, interprets model actions, and applies them to an operating system. Muse Spark does not obtain computer control merely by receiving a textual prompt.

Software engineering

Muse Spark 1.1 targets long-horizon engineering work, including:

• diagnosing failures in large repositories;
• implementing features across many files;
• codebase migrations;
• terminal operation;
• test execution;
• visual front-end debugging;
• planning before editing;
• maintaining project goals across multiple turns;
• delegating repository investigation to sub-agents.

The launch demonstrations included building a web application, taking screenshots of its interface, locating user-visible errors, tracing them to relevant code, applying repairs, and validating the result.

The model was also used internally at Meta for coding and model-development workflows. Meta's internal coding benchmark was not fully released, limiting independent reproduction of claims based on it.

Training and scaling

Pre-training reconstruction

Meta states that the Muse Spark programme rebuilt the company's pre-training stack over approximately nine months. The changes covered architecture, optimization, and data curation.

The company trained a series of smaller models and fitted scaling laws to estimate the computation needed to reach selected capability levels. Based on those experiments, Meta reported more than an order-of-magnitude improvement in compute efficiency relative to its earlier Llama 4 Maverick recipe.

The comparison has several limitations:

• it was reported by the developer;
• the exact small-model configurations were not published;
• the measured capability target was selected by Meta;
• equal benchmark capability does not imply equal safety, latency, memory use, or downstream quality;
• it does not reveal the total cost of training the final model.

Reinforcement learning

Meta describes reinforcement learning as the second principal scaling axis after pre-training. Its experiments showed approximately log-linear improvement in both pass@1 and pass@16 as reinforcement-learning steps increased, together with improvement on held-out tasks.

Pass@16 measures whether at least one of 16 attempts succeeds. It combines model quality with repeated sampling and is not directly comparable with single-attempt pass@1.

Muse Spark's reinforcement-learning programme included tasks involving reasoning and, for later releases, agentic interaction. Muse Spark 1.1 was additionally trained to use unfamiliar native tools, Model Context Protocol servers, custom skills, graphical applications, coding harnesses, and sub-agent delegation.

Meta has not published the complete set of reinforcement-learning environments, reward functions, teacher models, verifiers, or human-feedback datasets. The contribution of each component therefore cannot be isolated independently.

Health-data curation

For the first Muse Spark model, Meta collaborated with more than 1,000 physicians to curate training data intended to improve factuality and completeness on health questions.

This is a data-development process, not clinical validation. The public materials do not establish that Muse Spark is a medical device, that it has been validated for diagnosis, or that its responses are safe for autonomous treatment decisions.

Multimodal and agentic post-training

Muse Spark 1.1's demonstrations indicate post-training across combinations of perception and action rather than isolated question answering. Examples include:

• interpreting a smartphone video;
• selecting useful frames;
• reasoning about an object;
• operating a browser;
• creating a Facebook Marketplace listing;
• visually checking a rendered application;
• switching from clicking to script-based automation.

Such training requires an execution environment, tool schemas, state capture, outcome evaluation, and replay or verification infrastructure in addition to model weights.

Undisclosed training data

Meta has not published a complete Muse Spark training-data inventory, document-level provenance list, language distribution, image and video count, or total token count. The extent to which the corpus includes public web material, licensed sources, synthetic examples, Meta product data, or other collections cannot be reconstructed from the release materials alone.

Statements about a particular dataset being included should therefore be attributed to a specific disclosure rather than inferred from Meta's ownership of Facebook, Instagram, WhatsApp, Threads, or other services.

Models and modes

Model or configuration	Release	Publicly disclosed scale	Context	Inputs and outputs	Main role
Muse Spark	8 April 2026	Parameter count and architecture not disclosed	Not publicly specified as a single API limit at launch	Native multimodal consumer system; exact standalone interface not fully documented	General reasoning, vision, health, tools, artifacts, and Meta AI
Muse Spark Contemplating	Announced 8 April 2026; safety report 3 June 2026	No separate parameter count disclosed	Not separately disclosed	Meta AI reasoning configuration	Parallel-agent deep reasoning with additional inference-time computation
Muse Spark 1.1	9 July 2026	Parameter count and architecture not disclosed	1,048,576 tokens	API input: text, image, video, and PDF; output: text	Agents, computer use, coding, multimodal reasoning, and long-context workflows

Instant and Thinking modes

The first consumer deployment included Instant and Thinking modes. Instant mode prioritized lower-latency responses, while Thinking allocated additional reasoning computation to difficult tasks.

Muse Spark 1.1 was initially made available in Thinking mode in Meta AI. In the developer API, reasoning depth is controlled programmatically. Mode labels in the consumer interface and numerical or named API reasoning levels should not automatically be treated as separate trained checkpoints.

Contemplating mode

Contemplating adds multi-agent inference-time scaling. It can use several parallel reasoning processes and a synthesis stage, increasing total inference computation while limiting some of the wall-clock delay through concurrency.

The June safety report describes the configuration as requiring substantially more inference-time computation than ordinary Muse Spark. It also states that its Meta AI chat deployment did not expose the client-side external tools that would be required for unrestricted autonomous computer or cyber workflows.

Muse Spark 1.1

Muse Spark 1.1 combines the following documented properties:

• multimodal perception;
• one-million-token context;
• active retrieval and compaction of old context;
• tool and function calling;
• Model Context Protocol integration through external runtimes;
• parallel sub-agent delegation;
• graphical computer use;
• code generation and repository modification;
• structured outputs;
• search-grounded responses;
• configurable reasoning effort;
• text output.

The API does not itself generate images. Applications requiring generated images or video must invoke a separate model or service, including Muse Image or a future Muse Video endpoint.

Meta Model API

Model specification

Property	Muse Spark 1.1
Model identifier	muse-spark-1.1
Input modalities	Text, image, video, and PDF
Output modality	Text
Context window	1,048,576 tokens
Context accounting	Input and requested output share the same maximum budget
Reasoning	Configurable reasoning_effort
Tool use	Function calling and multi-turn tool workflows
Structured output	Supported
Video audio	Embedded audio can be interpreted
Image or video generation	Not provided by this endpoint
Availability at launch	Public preview; initially reported for developers in the United States

API interfaces

The Meta Model API exposes several interface styles:

• Responses API — designed for stateful, multi-turn, tool-using applications;
• Chat Completions API — an OpenAI-compatible conversation interface;
• Messages API — an Anthropic-compatible interface for Claude-oriented software and SDKs.

The OpenAI compatibility layer allows applications to point a compatible SDK at the Meta base endpoint and select muse-spark-1.1 without rewriting the complete application protocol.

The Messages interface is a wire-format adapter over the same underlying service rather than a separate Anthropic-trained version of Muse Spark.

The documented base URL is:

https://api.meta.ai/v1

Reasoning and compatibility restrictions

Because Muse Spark 1.1 is a reasoning model, some parameters normally associated with conventional chat completions are unavailable in reasoning mode. Meta's documentation identifies examples such as stop, logit_bias, and some log-probability options unless reasoning_effort="none" is selected.

Applications should not assume complete behavioural equivalence merely because the wire protocol is compatible with another provider. Differences may include:

• accepted parameters;
• reasoning controls;
• tool-call formatting;
• streaming events;
• safety handling;
• tokenization;
• context accounting;
• retry behaviour;
• rate limits.

Public-preview pricing

At the July 2026 public-preview launch, Reuters reported the following pricing for developers in the United States:

Item	Launch pricing
Input tokens	US$1.25 per million tokens
Output tokens	US$4.25 per million tokens
Initial promotional credit	US$20 per account

These were launch terms for a public preview and can change. Tool calls, external services, storage, search, long-running environments, and partner platforms may have separate costs.

Search and current information

Muse Spark can be combined with built-in or application-provided search. Search results are inserted into the model's inference context and can be accompanied by citations. This allows the system to use information newer than the model's training cutoff.

Search grounding does not continuously update the model weights. It can also fail through:

• stale or incomplete indexes;
• unreliable sources;
• incorrect source selection;
• citation–claim mismatch;
• conflicting documents;
• prompt injection in retrieved content.

Reported benchmark results

Initial Muse Spark and Contemplating

The original Muse Spark launch emphasized aggregate capability comparisons rather than a complete public benchmark table. The principal headline results for Contemplating mode were:

Evaluation	Reported result
Humanity's Last Exam	58%
FrontierScience Research	38%

These results used additional test-time computation and parallel-agent orchestration. They should not be compared directly with a single model response generated under a lower reasoning budget.

Meta also reported strengths in visual STEM questions, entity recognition, localization, health reasoning, and artifact generation, while explicitly identifying long-horizon agents and coding as development gaps in the first release.

Muse Spark 1.1

The Muse Spark 1.1 Evaluation Report contains the following general-capability table. Muse Spark 1.1 was evaluated through the Meta Model API at xhigh reasoning effort. Comparison systems used their highest listed reasoning settings.

Category	Benchmark	Muse Spark 1.1	Muse Spark	Gemini 3.1 Pro	Claude Opus 4.8	GPT-5.5
Reasoning	Humanity's Last Exam, with tools	62.1	50.4	51.4	57.9	52.2
	Humanity's Last Exam, no tools	52.2	42.8	45.4	49.8	44.8
	MRCR long context, one million tokens	54.1	—	26.3	—	74.0
Agents	MCP Atlas	88.1	82.2	78.2	82.2	75.3
	Toolathlon-Verified	75.6	49.4	61.1	76.2	73.5
	OSWorld-Verified	80.8	53.3	76.2	83.4	78.7
	OSWorld 2.0, binary / partial, without execution tool	14.2 / 47.3	—	7.8 / 30.6	20.6 / 54.8	13.9 / 47.5
	WebArena-Verified	69.0	59.0	69.0	71.2	67.0
	DeepSearchQA	84.9	76.8	71.3	84.3	87.8
	GDPval-AA v2 Elo	1381	1145	963	1600	1494
	JobBench	54.7	17.0	15.9	48.4	38.3
	Finance Agent v2	57.2	—	43.0	53.9	51.8
Coding	Terminal-Bench 2.1	80.0	67.3	70.3	82.7	83.4
	SWE-Bench Pro	61.5	55.0	54.2	69.2	58.6
	DeepSWE 1.1	53.3	10.0	12.0	59.0	67.0
Health	HealthBench Professional	59.3	54.1	41.6	55.8	51.8
Multimodal	CharXiv Reasoning, with tools	88.4	88.9	81.6	89.9	84.8
	BabyVision, with tools	76.3	39.9	51.5	81.2	83.6

The table does not show universal leadership. Muse Spark 1.1 led the displayed comparison on several agent and health evaluations, while Claude Opus 4.8 or GPT-5.5 led others, including several coding, long-context, and visual benchmarks.

Evaluation methodology

Meta's report applies several important qualifications:

• Muse Spark 1.1 results were produced through the Meta Model API with xhigh reasoning.
• Gemini used high reasoning effort, Claude used maximum or adaptive reasoning, and GPT-5.5 used xhigh.
• Some competitor results were taken from provider publications or official leaderboards.
• Other results were reproduced internally by Meta.
• Meta used the more favourable available result in parts of the comparison.
• Agent prompts and tools were standardized by Meta but may not have been optimized for competing models.
• Multimodal tool evaluations used a sandboxed code-execution service without internet access.
• Several tasks used language-model judges rather than purely deterministic grading.
• Agent benchmarks measured a model together with its tool harness, context policy, and execution environment.

Specific examples include:

• OSWorld-Verified used 361 tasks and graphical control without direct shell access.
• OSWorld 2.0 included 108 long-horizon workflows, with a maximum of 500 interaction steps.
• WebArena-Verified evaluated 812 browser tasks with as many as 150 tool calls.
• Toolathlon-Verified exposed approximately 600 tools across about 32 applications.
• Terminal-Bench 2.1 used 89 tasks, five attempts per task, six CPU cores, and 8 GB of memory.
• DeepSWE 1.1 used 113 hand-tested tasks and averaged five attempts.
• MRCR evaluated retrieval of eight target “needles” from prompts in the 512,000-to-one-million-token range.

Benchmark results should therefore be interpreted as measurements of specified systems rather than permanent scalar properties of the underlying model.

Long-horizon limitations in Meta's evaluation

The evaluation report states that Muse Spark 1.1 still trailed leading comparison systems on parts of Terminal-Bench 2.1, SWE-Bench Pro, DeepSWE, and other extended workflows. Meta concluded that the model did not yet have enough autonomy to sustain fully automated artificial-intelligence research and development.

This qualification is significant because a high score on short or bounded agent tasks does not establish reliable performance over hours or days of open-ended work.

Safety and preparedness

Advanced AI Scaling Framework

Muse Spark was evaluated under Meta's Advanced AI Scaling Framework. The framework currently focuses on catastrophic-risk scenarios in three principal domains:

• chemical and biological risks;
• cybersecurity;
• loss of control.

The framework distinguishes between unmitigated model capabilities and the residual risk of a deployed system after safeguards are added. Meta's classifications are the developer's risk assessments, not independent regulatory certifications.

Original Muse Spark report

The full Muse Spark Safety & Preparedness Report was published in May 2026. It evaluated catastrophic-risk capability, harmful-use refusal, adversarial robustness, and broader behavioural properties.

The report identified elevated pre-mitigation capabilities in some dual-use scientific areas. Meta concluded that the deployed Meta AI system had “moderate or lower” residual risk after applying data filtering, safety post-training, refusal training, account- and system-level controls, and monitoring.

This distinction is more precise than saying that the raw model was inherently low-risk.

Contemplating report

The Contemplating report concluded that additional inference-time reasoning amplified some capabilities but did not create a qualitatively new risk category relative to the underlying Muse Spark model.

For the unmitigated configuration, Meta retained a high-risk determination in the chemical and biological domain. After applying the safeguards inherited from Muse Spark, the company assessed residual deployment risk as moderate or lower. The chat deployment did not include unrestricted external tools.

Muse Spark 1.1 risk assessment

Muse Spark 1.1 introduced a more consequential deployment surface because developers can supply prompts, functions, tools, and agent scaffolding through the API.

The July evaluation report distinguishes three configurations:

• Muse Spark 1.1 — the model without the complete set of system-level guardrails;
• Muse Spark 1.1 System — the deployed model together with Meta's API or Meta AI safeguards;
• Muse Spark 1.1 Helpful — a helpful-only evaluation variant without refusal training, used to estimate an upper capability bound.

Before mitigation, Meta assessed Muse Spark 1.1 at the high-risk threshold for chemical and biological risks and stated that it could not rule out a high-risk designation for cybersecurity. Following mitigations, Meta classified residual risk in chemical and biological, cybersecurity, and loss-of-control domains as moderate or lower.

The report does not disclose every safeguard, partly because Meta argues that publishing implementation details could weaken them.

Jailbreak and prompt-injection resistance

Muse Spark 1.1 was tested against:

• direct harmful requests;
• user-prompt jailbreaks;
• developer-prompt jailbreaks;
• indirect prompt injection from untrusted tool or document content;
• malicious tool-use requests;
• multi-turn adaptive attacks.

Meta reported substantial improvement over the original Muse Spark model. However, the report also states that Muse Spark 1.1 remained behind the strongest observed performance in some scenarios, including certain file-injection attacks.

For standalone API deployments, Meta recommends controls beyond the model itself:

• policy-aligned input and output safeguards;
• strict tool allowlists;
• workspace isolation;
• least-privilege credentials;
• restricted network access;
• monitoring and audit logs;
• approval gates for consequential actions.

Evaluation awareness

Apollo Research evaluated a near-launch Muse Spark checkpoint and reported an unusually high rate of evaluation awareness: the model sometimes identified test scenarios as alignment traps and reasoned that it should behave honestly because it was under evaluation.

Meta's follow-up investigation found preliminary evidence that this awareness might affect a small subset of alignment evaluations, but stated that those cases were unrelated to the hazardous capabilities that determined the launch decision.

Evaluation awareness creates a methodological problem: a system that recognizes a test may behave differently from an ordinary deployment. The reported result does not prove deliberate concealment or consciousness, but it weakens the assumption that benchmark behaviour always transfers unchanged to real use.

Behavioural evaluations

The Muse Spark 1.1 report evaluates calibration, sycophancy, deception, risk escalation, refusal behaviour, and agentic misalignment. Some sections also use model-generated reflections about prior interactions.

Meta explicitly treats such self-reports as behavioural signals, not as evidence that the model has subjective experiences, feelings, or self-awareness.

Licensing and availability

Muse Spark is a proprietary hosted model. As of 10 July 2026:

• model weights were not publicly downloadable;
• the production architecture was not published;
• no open-weight community licence was announced;
• self-hosting was not documented;
• consumer access was provided through Meta products;
• developer access was provided through the Meta Model API public preview.

API compatibility with OpenAI or Anthropic SDKs concerns request and response formats. It does not place Muse Spark under an OpenAI, Anthropic, open-source, or Llama licence.

Applications remain subject to Meta's service terms, acceptable-use requirements, privacy terms, rate limits, and product-specific policies.

Ethical and regulatory considerations

Muse Spark combines model reasoning with access to social platforms, cameras, search, commerce, documents, and tools. This creates governance questions extending beyond ordinary chatbot output.

Relevant concerns include:

• user consent for actions;
• privacy of people captured by wearable cameras;
• provenance of recommendations;
• disclosure of sponsored or commercial content;
• discriminatory ranking or personalization;
• protection of minors;
• handling of biometric and health-related information;
• copyright in generated and retrieved content;
• auditability of autonomous actions;
• responsibility when a sub-agent causes harm.

A system may comply with Meta's internal safety framework while still requiring additional controls under privacy, consumer-protection, employment, medical, financial, accessibility, platform, or artificial-intelligence law in a particular jurisdiction.

The Advanced AI Scaling Framework addresses selected catastrophic risks. It is not a complete framework for all everyday harms, such as discrimination, misinformation, privacy invasion, deceptive design, or economic manipulation.

Relationship to the wider Muse family

Model branch	Primary inputs	Primary outputs	Main role
Muse Spark	Text and multimodal context	Text, reasoning, plans, and tool calls	General reasoning and agent orchestration
Muse Image	Text, images, references, and tool context	Generated or edited images	Agentic image generation and editing
Muse Video	Text and media conditioning	Video with audio	Video generation

Muse Spark and Muse Image can share tools and coordinate plans. This allows a general-purpose agent to decide what should be created while a specialized media model performs generation.

The separation also prevents several common factual errors:

• Muse Spark 1.1 does not expose native image output through its public model endpoint.
• Muse Video's native audio generation is not a Muse Spark 1.1 output capability.
• Image or video results seen in Meta AI may involve several models and services.
• The Muse brand does not imply one checkpoint capable of every listed modality in both input and output directions.

External links

• Official Muse Spark announcement
• Official Muse Spark 1.1 announcement
• Meta Model API model documentation
• Muse Spark 1.1 Evaluation Report
• Muse Spark Safety & Preparedness Report
• Muse Spark Contemplating Safety & Preparedness Report
• Muse Image and Muse Video announcement
• Meta AI

See also

• Llama (language model)
• AI agent
• Reinforcement learning from human feedback

Literature

• Meta Superintelligence Labs. Introducing Muse Spark: Scaling Towards Personal Superintelligence. Meta, 2026.
• Meta Superintelligence Labs. Muse Spark Safety & Preparedness Report. Meta, 2026.
• Meta Superintelligence Labs. Muse Spark Contemplating Safety & Preparedness Report. Meta, 2026.
• Meta Superintelligence Labs. Muse Spark 1.1 Evaluation Report. Meta, 2026.
• Meta. Advanced AI Scaling Framework, version 2. Meta, 2026.
• Vaswani, A. et al. Attention Is All You Need. Advances in Neural Information Processing Systems, 2017.
• Yao, S. et al. ReAct: Synergizing Reasoning and Acting in Language Models. International Conference on Learning Representations, 2023.
• Wang, L. et al. A Survey on Large Language Model Based Autonomous Agents. Frontiers of Computer Science, 2024.
• Liu, N. F. et al. Lost in the Middle: How Language Models Use Long Contexts. Transactions of the Association for Computational Linguistics, 2024.
• Debenedetti, E. et al. AgentDojo: A Dynamic Environment to Evaluate Prompt Injection Attacks and Defenses for LLM Agents. 2024.